Research On Key Techniques Of Image Perceptual Hashing

It is well known that the multimedia era has come upon us. Because images areperfect for conveying large information directly and vividly, they have become a mediaas important as texts. With the increasingly important role in the modern society, higherrequirements for image management and security insurances have arisen. Through thebrief digesting of the content of an image and the digest based matching, image perceptu-al hashing is a promising technique to meet the multiple application requirements and hasdrawn many attentions. However, the theoretical foundation of perceptual hashing is notmature, the understanding to the performance requirements and evaluation in different ap-plication modes have not been clarified, and multiple performances of existing perceptualhashing algorithms cannot meet the applicable requirements. So it is necessary to conquerthe key tequchiques of perceptual hashing, provide the theoritcial support, design the newalgorithms with better performance and improve the healthy and sustainable developmentof perceptual hashing and the related indurstry.After studying the basic concepts, application modes, state-of-arts of perceptualhashing systematically, current problems of perceptual hashing technique are summa-rized. Aiming at solving these problems, the creative contributions of this dissertationinclude the follows.The main performances and their evaluation methods in recognition mode and au-thentication mode are studied systematically. In recognition mode, current shortage andimprovement direction are pointed out via the analysis on the inner-test and inter-testbased robustness/discrimination test methods. In authentication mode, the concept of ver-ifiability and the Identical Perceptual Distance (IPD) based test method is presented toevaluate the ability of one perceptual hashing algorithm to differentiate the local tamperand global content preserving operations. The concepts of onewayness, confusion, dif-fusion and collision-resistance are compared and studied, and the relationships betweenthese features and security are clarified. For diffusion, which has no test method so far,the Flip Rate Distribution (FRD) curve based test method is proposed. The experimen-tal results demonstrate the effectiveness of the proposed IPD curve and FRD curve testmethods. Aiming at the inefficient robustness to geometric attack in recognition mode, theScale Invariant Feature Transform (SIFT) feature point based image perceptual hashingalgorithm is proposed. Through the filter, compression and centroid of feature pointsbased quantization, the perceptual digest is generated. The matching algorithm is based ongeneral set distance. The algorithm can preserve the identifying ability of the SIFT featurepoint to the image well, and resist common content preserving operations, especiallygeometric attacks such as rotation, scale and cropping.To improve the tamper detection ability in authentication mode, two perceptual hash-ing algorithms are proposed, which are based information fusion and saliency map re-spectively. The information fusion algorithm utilizes the advantages of different types offeatures. The global feature is responsible for the robustness/discrimination, and the localfeature is in charge of tamper detection. The saliency map based algorithm extracts theimage salient feature compatible with human neurophysiology and cognitive psycholo-gy, which makes it possible to detect the meaningful tamper. The experimental resultson a middle scale image database demonstrate that the proposed algorithms implementmuch higher tamper detection ability than existing algorithms except for the comparativerobustness/discrimination.The common problem of low diffusion in current perceptual hashing algorithms de-creases the security directly. To overcome such problem, a fuzzy commitment based per-ceptual hashing authentication scheme is proposed. The analysis and experiment resultsshow that the scheme can improve the security greatly without sacrificing the robustnessand discrimination of original perceptual hashing algorithm. The security problem due tolow diffusion can also be solved.